Compilation performance comparison of 3.5.0 and TreeSSA trees on MICO sources as requested in: [tree-ssa] Merge status 2004-05-03

[Matt Austern]

On May 5, 2004, at 11:49 AM, Laurent GUERBY wrote:

> On Wed, 2004-05-05 at 20:35, Joe Buck wrote:
>> The purpose of doing some optimization is to make -O0 *faster*.  We've
>> already got a 20% speed degradation any you want to allow 5% *more*?
>
> You want to allow some room IMHO:
>
> - there will always be some cases where -Og is slower than -O0
> and we don't want endless case by case discussions
>
> - If the generated file size is reduced by 50% for 10% longer compile
> time, I'd like to get such things in.
>
> Nobody will notice 44 minutes vs 40 minutes of full app compilation,
> everyone in a edit/compile one file/link full app/debug cycle will
> notice immediately a 50% reduction in link time.

I'm with Joe.  We're in the state we're in now because too many
people have said that "nobody will notice" yet another performance
degradation.  The argument that nobody will notice a small
regression is especially dangerous because it's true: nobody does
ever notice a 1% regression.  But if we let in dozens of 1%
regressions (which we have), then people start noticing.

The immediate question under discussion is: how do we fix a 20%
speed regression.  This is urgent.  We don't fix a 20% regression by
creating a new compilation mode that adds a 10% regression on
top of it.

			--Matt